alantrrs / opentld Goto Github PK

The function implemented to obtain the variance using integral images is not working properly.

 ./run_tld ../parameters.yml -s motorbike.mp4 

Source video (motorbike.mp4) appears misaligned. Its slanted diagonally across the screen. This is weird because it plays correctly in other applications.

Either a couple of pixels are added at the end of each line, shifting it across a little too much each time. Or otherwise the horizontal width of the window is not quite enough. The window seems too small by 2 pixel per line.

I'm not sure if its OpenTLD or the input video. It would be a good idea for others to try using different source video also. With the new -s switch

I like your work, it is so cool. I think the function "filterPts" don't need to
check "status" twice and the second check is wrong.
And I find the function TLD::getOverlappingBoxes is not efficient.the nth_element waste much time, and the waste is not necessary.
Best wish~

I an using ubuntu13.10 with
linux kernel 3.11.0-14
gcc 4.8.1
opencv 2.4.5
cmake 2.8.11.2

I add FindOpenCV.cmake
set (OpenCV_DIR /home/Documents/OpenCV/OpenTLD/build)
to the src folder and run cmake, everything goes well.
but when I run make, it comes up with many link error
undefined reference to 'cv::Mat......

So, what is wrong?

Alan,

I've downloaded and had your code working before, but I just redownloaded and built it, everything went smooth, but when I run it now based on the command given in the README, it seg faults when I select the bounding box. In fact, the very first time I ran it, it froze every time i moved the window or tried specifying the bounding box. Now it just seg faults everytime

Hi,

I'm not sure if others will see this or not. It happened to me whilst dragging the white rectangle the 3rd or 4th time.

janitors-Mac-mini:bin id$ ./run_tld parameter.yml -s motorbike.mp4 
Bounding box too small, try again.
Bounding box too small, try again.
Bounding box too small, try again.
Created 819074 bounding boxes
Found 0 good boxes, 0 bad boxes
Best Box: 13 69 61 41
Bounding box hull: 2147483647 2147483647 -2147483647 -2147483647
OpenCV Error: Assertion failed (!(dsize == Size()) || (inv_scale_x > 0 && inv_scale_y > 0)) in resize, file /tmp/homebrew-opencv-2.2-hzav/OpenCV-2.2.0/modules/imgproc/src/imgwarp.cpp, line 1392
terminate called after throwing an instance of 'cv::Exception'
  what():  /tmp/homebrew-opencv-2.2-hzav/OpenCV-2.2.0/modules/imgproc/src/imgwarp.cpp:1392: error: (-215) !(dsize == Size()) || (inv_scale_x > 0 && inv_scale_y > 0) in function resize

Abort trap

Thanks for the hard work on porting the matlab code over to C++. I intended to port your code to ccv: https://github.com/liuliu/ccv, but find the licence is incompatible, since it is a C++ code without any original matlab code, you can have an alternative licence other than the author's original GPL3, can you add a more liberal licence to your project such as MIT, Apache or FreeBSD licence? Thanks!

Hi Allan,

I'm using Qt 4.3.7,OpenCV 2.3,MS VS2010

I have a exception in the conversion of the 8Bit B/W-Image to the 32 bit unsigned cvMat.
bool Detector::init(IplImage& image, const QRectF& region)
{
…
cv::Mat_ imageMat(m_image); <--- crashes wenn converting the 8 Bit BW-Image to 32 Bit cvMat
because there is no conversation from CV_8U to CV_32U
in modules\core\src\convert.cpp/cv::Mat::convertTo()

cv::Mat_<int> imageMat(m_image);    works, because there is a conversation from CV_8U to CV_32U

}
What is wrong with that? I saw another Thread with someone using also OpenCV2.3, without this Problem.

Cheers
Rudi

Detector make too much false positive and misses a lot of detections. I believe the ensemble classifier is the problem.

tracker.filterPts()
It should be written this way:

//before using this function,points whose status is false has already been filted.
bool LKTracker::filterPts(vec_POINT2F *pts1,vec_POINT2F *pts2)
{
short i,a;
short n = pts2->size;

//计算FBerror和ncc的中值
medNCC = median(&similarity);
medFB = median(&FB_error);

//过滤掉相似度较低的那一半
for( i=a=0; i<n; ++i )
{
    if( vec_at(similarity,i) < medNCC || vec_at(FB_error,i) > medFB )
        continue;

    vec_at(*pts1,a) = vec_at(*pts1,i);
    vec_at(*pts2,a) = vec_at(*pts2,i);
    vec_at(FB_error,a) = vec_at(FB_error,i);
    a++;
}

vec_POINT2F_resize(pts1,a);
vec_POINT2F_resize(pts2,a);

if (a>0)
    return true;
else
    return false;

}

Check ZK's code,then you will know I'm right

Error 52 error LNK1104: cannot open file 'C:\Users\Lauren\Downloads\OpenTLD-master\OpenTLD-master\lib\Debug\tld.lib' C:\blah2\LINK
This error appears when compiling run_tld. The file tld.lib is missing? Any known solutions? Running cmake generates 4 .lib files but tld.lib is not generated

I am wordering why "num_closest_init" never get a value.
In my tests in get 32 good boxes but it is resized to zero in "TLD::init" and bbhull get abnormal values.

I don't understand the code:
matchTemplate(m_pEx[i], example, ncc, CV_TM_CCORR_NORMED); // measure NCC to positive examples
nccP = (ncc.at(0) + 1) * 0.5f;
I think the nccP's result can't > 0.95 forever.
who can tell me the writer‘s mean?
Thanks very much.

I'm running ubuntu 11.10 on a 32-bit system, practically vanilla with opencv 2.0. The build system refuses to compile. The code I'm using is at my personal repository, unchanged.

It runs very slow on my laptop. I use a 640x480 webcam.

""
640x480 webcam
Ubuntu 10.10
OpenCV 2.3
2.4 GHz Dual Core
4G RAM
""

I use the default parameter in the YAML file.
Here is the output from the console:

""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""
Created 614610 bounding boxes
Found 10 good boxes, 613623 bad boxes
Best Box: 417 225 80 82
Bounding box hull: 409 211 98 108
var: 1080.19
Positive fern examples generated: 200
negative data generation started.
Negative examples generated: 7357
Fern detector made 19 detections in 7749.08ms
Found 5 NN matches
Tracking...
Cluster indexes: 0
Found 5 clusters
No confident cluster was found
Low variance..not training
""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

Is there any parameter I should tune to make it faster?

I an using ubuntu14.04 with
linux kernel 3.13.0-24-generic
gcc 4.8.2
opencv 2.4.6
cmake 2.8.12.2

Everything is find, but at running the command "sudo ./run_tld -p ../paramers.yml -s ../datasets/06_car/car.mpg" ,The program crashed.

Error message is following***********
Initial Bounding Box = x:131 y:116 h:114 w:55
Created 269028 bounding boxes
Found 0 good boxes, 0 bad boxes
Best Box: 133 115 114 55
Bounding box hull: 2147483647 2147483647 -2147483647 -2147483647
OpenCV Error: Assertion failed (dsize.area() || (inv_scale_x > 0 && inv_scale_y > 0)) in resize, file /home/johnqi/softWare/opencv-2.4.6.1/modules/imgproc/src/imgwarp.cpp, line 1724
terminate called after throwing an instance of 'cv::Exception'
what(): /home/johnqi/softWare/opencv-2.4.6.1/modules/imgproc/src/imgwarp.cpp:1724: error: (-215) dsize.area() || (inv_scale_x > 0 && inv_scale_y > 0) in function resize

I doesn't change the program at all, So, what is wrong?
Thanks, in advance.

I find that tracking will almost always start, and will also often continue into the video for at least a little while. But segfaults also often occur. For example:

https://gist.github.com/1045500

After maybe about 10 runs I found that the tracker continued persistenly (without segfault) on about 2-3 occasions. One time with a continued tracking rectangle / clusterpoints, and one time it continued with lost tracking. All the others were segfaults.

Tracking happens but is hard for me to give feedback there, because of the other problem where my video is shifted across. Anyway heres some screenshots. They were taken from the 1 successful tracking session:

https://gist.github.com/1045534

https://gist.github.com/1045535

https://gist.github.com/1045536